stats

type Dimension string

type Filters struct {
	Providers    []string
	Runs         []string
	Models       []string
	Suites       []string
	Categories   []string
	Difficulties []string
	Statuses     []string
	Tags         []string
	TagMode      TagMode
}

type OutputFormat string

type Record struct {
	// Dimensions maps each requested group-by dimension name to this record's value for it
	// (e.g. {"provider": "openai", "run": "gpt-4"}).
	Dimensions map[string]string

	// Count is the number of non-skipped results (Passed + Failed + Error).
	Count   int
	Passed  int
	Failed  int
	Error   int
	Skipped int

	// PassRate, Accuracy and ErrorRate are percentages (0-100) rounded to 2 decimals. See
	// formatters.PassRate/AccuracyRate/ErrorRate for the exact formulas; a zero denominator
	// yields 0, not an undefined value.
	PassRate  float64
	Accuracy  float64
	ErrorRate float64

	// Duration metrics are computed over non-skipped results only.
	TotalDuration  *time.Duration
	MedianDuration *time.Duration
	StddevDuration *time.Duration

	// Token/tool-call metrics reflect only the candidate answer and any subsequent error
	// (never judge/validation usage), matching the HTML report's dynamic summary. A metric
	// is nil when no contributing result reported it.
	//
	// Input metrics resolve cache tokens according to InputTokenAccounting (see
	// TokenUsage.EffectiveInputTokens), so they are comparable across providers regardless
	// of whether cache reads/writes are already part of InputTokens.
	TotalInputTokens  *int64
	MedianInputTokens *float64
	StddevInputTokens *float64

	// Output metrics resolve reasoning tokens according to OutputTokenAccounting (see
	// TokenUsage.GeneratedTokens), so they are comparable across providers regardless of
	// whether reasoning is already part of OutputTokens.
	TotalOutputTokens  *int64
	MedianOutputTokens *float64
	StddevOutputTokens *float64

	// Reasoning/cache metrics sum only the counts providers actually reported, so they stay
	// nil when none did rather than implying zero.
	TotalReasoningTokens  *int64
	MedianReasoningTokens *float64
	StddevReasoningTokens *float64

	TotalCacheReadTokens  *int64
	MedianCacheReadTokens *float64
	StddevCacheReadTokens *float64

	TotalCacheWriteTokens  *int64
	MedianCacheWriteTokens *float64
	StddevCacheWriteTokens *float64

	TotalToolCalls  *int64
	MedianToolCalls *float64
	StddevToolCalls *float64

	// EstimatedCandidateCost derives from the static prices the candidate run was configured
	// with and is never a billed amount. It is nil when any contributing result reported
	// usage that could not be priced, so a partial sum is never mistaken for a complete total.
	EstimatedCandidateCost *float64

	// CandidateCostCurrency is the ISO 4217 code EstimatedCandidateCost is expressed in,
	// empty when unknown.
	CandidateCostCurrency string

	// TransientErrors and ResponseParsingErrors count Error-kind results whose
	// ErrorDetails.Transient/ResponseParsing flag is explicitly true. The two are not
	// mutually exclusive: a result can set both.
	TransientErrors       int
	ResponseParsingErrors int
}

type TagMode string